[NIXL][PCP][DCP] Expose PCP producer KV shards as transfer ranks - #3
Closed
LucasWilkinson wants to merge 1 commit into
Closed
LucasWilkinson wants to merge 1 commit into
LucasWilkinson wants to merge 1 commit into
Conversation
Rebase vllm-project#54496 onto vllm-project#56157. Publish each TP1 PCP+DCP producer shard through the existing DCP transfer topology and retain upstream completion aggregation. Share transfer-size calculation between the pull and push schedulers and validate the PCP/DCP group in the connector. Cover shard publication, completion, push scheduling, and advertised transfer topology for replicated and sharded producers. Co-authored-by: OpenAI Codex <noreply@openai.com> Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Author
|
Recreated against vllm-project/vllm:main now that vllm-project#56157 has merged: vllm-project#56645 |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Expose TP1 PCP+DCP KV shards as NIXL transfer ranks, reusing DCP mapping and completion handling.
Stacks on vllm-project#56157; replaces vllm-project#54496. Adds transfer topology missing from vllm-project#56157; distinct from vllm-project#45340's block accounting.
GLM-5.2-NVFP4, PCP4+TP1+EP4+DCP4 → TP4, with additional accuracy/offload fixes: controlled GSM8K 28/32 with GPU caching and 28/32 with CPU KV offload (direct TP4: 26/32). All 32 GPU/CPU pairs matched tokens and log probabilities exactly, with verified CPU hits. Long-context CPU restores passed; NIXL byte comparison, including a partial block: 146,716,800 bytes, zero mismatches. These model results require the linked fixes; they are not results for this PR alone.
Validation: 89 focused NIXL tests passed; additional-fix suite 502 passed, 19 skipped, followed by 101 passing indexer tests after final edits. Pre-commit/mypy passed.
Test commands
.venv/bin/python -m pytest tests/v1/kv_connector/unit/test_nixl_connector.py -k 'pcp_producer or scheduler_advertises_transfer_topology' .venv/bin/python -m pytest tests/v1/kv_connector/unit/{test_tp_mapping,test_nixl_push_connector}.py .venv/bin/python -m pytest tests/v1/attention/{test_indexer_dcp_localize,test_flashmla_sparse_pcp_dcp}.py tests/v1/simple_kv_offload/{test_worker,test_scheduler}.py tests/kernels/test_top_k_per_row.py -q .venv/bin/python -m pytest tests/v1/attention/test_indexer_dcp_localize.py -qAI assistance: OpenAI Codex. Draft pending human review and validation.